Fio 1.27.1
[fio.git] / init.c
diff --git a/init.c b/init.c
index 4f57ba31705e151f44859ce80109d3fb1c3a110e..2e7a0fe38854ace4c0a4a90e9fa575d2c7013e92 100644 (file)
--- a/init.c
+++ b/init.c
 
 #include "fio.h"
 #include "parse.h"
+#include "smalloc.h"
+#include "filehash.h"
+#include "verify.h"
 
-static char fio_version_string[] = "fio 1.19";
+static char fio_version_string[] = "fio 1.27.1";
 
 #define FIO_RANDSEED           (0xb1899bedUL)
 
@@ -25,7 +28,7 @@ static char **ini_file;
 static int max_jobs = MAX_JOBS;
 static int dump_cmdline;
 
-struct thread_data def_thread;
+static struct thread_data def_thread;
 struct thread_data *threads = NULL;
 
 int exitall_on_terminate = 0;
@@ -39,18 +42,20 @@ char *job_section = NULL;
 int write_bw_log = 0;
 int read_only = 0;
 
-static int def_timeout = 0;
-static int write_lat_log = 0;
+static int def_timeout;
+static int write_lat_log;
 
 static int prev_group_jobs;
 
 unsigned long fio_debug = 0;
+unsigned int fio_debug_jobno = -1;
+unsigned int *fio_debug_jobp = NULL;
 
 /*
  * Command line options. These will contain the above, plus a few
  * extra that only pertain to fio itself and not jobs.
  */
-static struct option long_options[FIO_NR_OPTIONS] = {
+static struct option l_opts[FIO_NR_OPTIONS] = {
        {
                .name           = "output",
                .has_arg        = required_argument,
@@ -116,6 +121,11 @@ static struct option long_options[FIO_NR_OPTIONS] = {
                .has_arg        = required_argument,
                .val            = 'x',
        },
+       {
+               .name           = "alloc-size",
+               .has_arg        = required_argument,
+               .val            = 'a',
+       },
        {
                .name           = NULL,
        },
@@ -197,6 +207,13 @@ static int setup_rate(struct thread_data *td)
        return 0;
 }
 
+static int fixed_block_size(struct thread_options *o)
+{
+       return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
+               o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
+               o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
+}
+
 /*
  * Lazy way of fixing up options that depend on each other. We could also
  * define option callback handlers, but this is easier.
@@ -205,13 +222,28 @@ static int fixup_options(struct thread_data *td)
 {
        struct thread_options *o = &td->o;
 
+#ifndef FIO_HAVE_PSHARED_MUTEX
+       if (!td->o.use_thread) {
+               log_info("fio: this platform does not support process shared"
+                        " mutexes, forcing use of threads. Use the 'thread'"
+                        " option to get rid of this warning.\n");
+               td->o.use_thread = 1;
+       }
+#endif
+
+#ifndef FIO_HAVE_CPU_AFFINITY
+       if (td->o.gtod_cpu) {
+               log_err("fio: platform must support CPU affinity for"
+                       "gettimeofday() offloading\n");
+               return 1;
+       }
+#endif
+
        if (read_only && td_write(td)) {
-               log_err("fio: job <%s> has write bit set, but fio is in read-only mode\n", td->o.name);
+               log_err("fio: job <%s> has write bit set, but fio is in"
+                       " read-only mode\n", td->o.name);
                return 1;
        }
-       
-       if (o->rwmix[DDIR_READ] + o->rwmix[DDIR_WRITE] > 100)
-               o->rwmix[DDIR_WRITE] = 100 - o->rwmix[DDIR_READ];
 
        if (o->write_iolog_file && o->read_iolog_file) {
                log_err("fio: read iolog overrides write_iolog\n");
@@ -232,21 +264,38 @@ static int fixup_options(struct thread_data *td)
                o->overwrite = 1;
 
        if (!o->min_bs[DDIR_READ])
-               o->min_bs[DDIR_READ]= o->bs[DDIR_READ];
+               o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
        if (!o->max_bs[DDIR_READ])
                o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
        if (!o->min_bs[DDIR_WRITE])
-               o->min_bs[DDIR_WRITE]= o->bs[DDIR_WRITE];
+               o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
        if (!o->max_bs[DDIR_WRITE])
                o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
 
        o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
 
+       /*
+        * For random IO, allow blockalign offset other than min_bs.
+        */
+       if (!o->ba[DDIR_READ] || !td_random(td))
+               o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
+       if (!o->ba[DDIR_WRITE] || !td_random(td))
+               o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
+
+       if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
+           o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
+           !td->o.norandommap) {
+               log_err("fio: Any use of blockalign= turns off randommap\n");
+               td->o.norandommap = 1;
+       }
+
        if (!o->file_size_high)
                o->file_size_high = o->file_size_low;
 
-       if (o->norandommap && o->verify != VERIFY_NONE) {
-               log_err("fio: norandommap given, verify disabled\n");
+       if (o->norandommap && o->verify != VERIFY_NONE
+           && !fixed_block_size(o))  {
+               log_err("fio: norandommap given for variable block sizes, "
+                       "verify disabled\n");
                o->verify = VERIFY_NONE;
        }
        if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
@@ -300,12 +349,18 @@ static int fixup_options(struct thread_data *td)
        }
 
        if (o->fill_device && !o->size)
-               o->size = ULONG_LONG_MAX;
-       
+               o->size = -1ULL;
+
        if (td_rw(td) && td->o.verify != VERIFY_NONE)
                log_info("fio: mixed read/write workload with verify. May not "
                 "work as expected, unless you pre-populated the file\n");
 
+       if (td->o.verify != VERIFY_NONE)
+               td->o.refill_buffers = 1;
+
+       if (td->o.pre_read)
+               td->o.invalidate_cache = 0;
+
        return 0;
 }
 
@@ -357,13 +412,32 @@ static int exists_and_not_file(const char *filename)
        return 1;
 }
 
+void td_fill_rand_seeds(struct thread_data *td)
+{
+       os_random_seed(td->rand_seeds[0], &td->bsrange_state);
+       os_random_seed(td->rand_seeds[1], &td->verify_state);
+       os_random_seed(td->rand_seeds[2], &td->rwmix_state);
+
+       if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
+               os_random_seed(td->rand_seeds[3], &td->next_file_state);
+
+       os_random_seed(td->rand_seeds[5], &td->file_size_state);
+
+       if (!td_random(td))
+               return;
+
+       if (td->o.rand_repeatable)
+               td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
+
+       os_random_seed(td->rand_seeds[4], &td->random_state);
+}
+
 /*
  * Initialize the various random states we need (random io, block size ranges,
  * read/write mix, etc).
  */
 static int init_random_state(struct thread_data *td)
 {
-       unsigned long seeds[6];
        int fd;
 
        fd = open("/dev/urandom", O_RDONLY);
@@ -372,30 +446,15 @@ static int init_random_state(struct thread_data *td)
                return 1;
        }
 
-       if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
+       if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
+           (int) sizeof(td->rand_seeds)) {
                td_verror(td, EIO, "read");
                close(fd);
                return 1;
        }
 
        close(fd);
-
-       os_random_seed(seeds[0], &td->bsrange_state);
-       os_random_seed(seeds[1], &td->verify_state);
-       os_random_seed(seeds[2], &td->rwmix_state);
-
-       if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
-               os_random_seed(seeds[3], &td->next_file_state);
-
-       os_random_seed(seeds[5], &td->file_size_state);
-
-       if (!td_random(td))
-               return 0;
-
-       if (td->o.rand_repeatable)
-               seeds[4] = FIO_RANDSEED * td->thread_number;
-
-       os_random_seed(seeds[4], &td->random_state);
+       td_fill_rand_seeds(td);
        return 0;
 }
 
@@ -443,14 +502,15 @@ static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
                td->io_ops->flags |= FIO_RAWIO;
 
        file_alloced = 0;
-       if (!td->o.filename && !td->files_index) {
+       if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
                file_alloced = 1;
 
                if (td->o.nr_files == 1 && exists_and_not_file(jobname))
                        add_file(td, jobname);
                else {
                        for (i = 0; i < td->o.nr_files; i++) {
-                               sprintf(fname, "%s.%d.%d", jobname, td->thread_number, i);
+                               sprintf(fname, "%s.%d.%d", jobname,
+                                                       td->thread_number, i);
                                add_file(td, fname);
                        }
                }
@@ -466,7 +526,7 @@ static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
                        f->real_file_size = -1ULL;
        }
 
-       td->mutex = fio_sem_init(0);
+       td->mutex = fio_mutex_init(0);
 
        td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
        td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
@@ -500,9 +560,12 @@ static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 
        if (!terse_output) {
                if (!job_add_num) {
-                       if (!strcmp(td->io_ops->name, "cpuio"))
-                               log_info("%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->o.name, td->o.cpuload, td->o.cpucycle);
-                       else {
+                       if (!strcmp(td->io_ops->name, "cpuio")) {
+                               log_info("%s: ioengine=cpu, cpuload=%u,"
+                                        " cpucycle=%u\n", td->o.name,
+                                                       td->o.cpuload,
+                                                       td->o.cpucycle);
+                       } else {
                                char *c1, *c2, *c3, *c4;
 
                                c1 = to_kmg(td->o.min_bs[DDIR_READ]);
@@ -510,7 +573,13 @@ static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
                                c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
                                c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
 
-                               log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s, ioengine=%s, iodepth=%u\n", td->o.name, td->groupid, ddir_str[td->o.td_ddir], c1, c2, c3, c4, td->io_ops->name, td->o.iodepth);
+                               log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
+                                        " ioengine=%s, iodepth=%u\n",
+                                               td->o.name, td->groupid,
+                                               ddir_str[td->o.td_ddir],
+                                               c1, c2, c3, c4,
+                                               td->io_ops->name,
+                                               td->o.iodepth);
 
                                free(c1);
                                free(c2);
@@ -539,6 +608,7 @@ static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
                if (file_alloced) {
                        td_new->o.filename = NULL;
                        td_new->files_index = 0;
+                       td_new->files_size = 0;
                        td_new->files = NULL;
                }
 
@@ -561,7 +631,7 @@ static int skip_this_section(const char *name)
        if (!strncmp(name, "global", 6))
                return 0;
 
-       return strncmp(job_section, name, strlen(job_section));
+       return strcmp(job_section, name);
 }
 
 static int is_empty_or_comment(char *line)
@@ -594,6 +664,8 @@ static int parse_jobs_ini(char *file, int stonewall_flag)
        int first_sect = 1;
        int skip_fgets = 0;
        int inside_skip = 0;
+       char **opts;
+       int i, alloc_opts, num_opts;
 
        if (!strcmp(file, "-"))
                f = stdin;
@@ -613,6 +685,10 @@ static int parse_jobs_ini(char *file, int stonewall_flag)
        name = malloc(280);
        memset(name, 0, 280);
 
+       alloc_opts = 8;
+       opts = malloc(sizeof(char *) * alloc_opts);
+       num_opts = 0;
+
        stonewall = stonewall_flag;
        do {
                /*
@@ -634,7 +710,8 @@ static int parse_jobs_ini(char *file, int stonewall_flag)
                if (sscanf(p, "[%255s]", name) != 1) {
                        if (inside_skip)
                                continue;
-                       log_err("fio: option <%s> outside of [] job section\n", p);
+                       log_err("fio: option <%s> outside of [] job section\n",
+                                                                       p);
                        break;
                }
 
@@ -670,6 +747,9 @@ static int parse_jobs_ini(char *file, int stonewall_flag)
                        stonewall = 0;
                }
 
+               num_opts = 0;
+               memset(opts, 0, alloc_opts * sizeof(char *));
+
                while ((p = fgets(string, 4096, f)) != NULL) {
                        if (is_empty_or_comment(p))
                                continue;
@@ -687,29 +767,42 @@ static int parse_jobs_ini(char *file, int stonewall_flag)
 
                        strip_blank_end(p);
 
-                       /*
-                        * Don't break here, continue parsing options so we
-                        * dump all the bad ones. Makes trial/error fixups
-                        * easier on the user.
-                        */
-                       ret |= fio_option_parse(td, p);
-                       if (!ret && dump_cmdline)
-                               log_info("--%s ", p);
+                       if (num_opts == alloc_opts) {
+                               alloc_opts <<= 1;
+                               opts = realloc(opts,
+                                               alloc_opts * sizeof(char *));
+                       }
+
+                       opts[num_opts] = strdup(p);
+                       num_opts++;
                }
 
-               if (!ret)
+               ret = fio_options_parse(td, opts, num_opts);
+               if (!ret) {
+                       if (dump_cmdline)
+                               for (i = 0; i < num_opts; i++)
+                                       log_info("--%s ", opts[i]);
+
                        ret = add_job(td, name, 0);
-               else {
+               else {
                        log_err("fio: job %s dropped\n", name);
                        put_job(td);
                }
+
+               for (i = 0; i < num_opts; i++)
+                       free(opts[i]);
+               num_opts = 0;
        } while (!ret);
 
        if (dump_cmdline)
                log_info("\n");
 
+       for (i = 0; i < num_opts; i++)
+               free(opts[i]);
+
        free(string);
        free(name);
+       free(opts);
        if (f != stdin)
                fclose(f);
        return ret;
@@ -738,19 +831,28 @@ static void free_shm(void)
        struct shmid_ds sbuf;
 
        if (threads) {
-               shmdt((void *) threads);
+               void *tp = threads;
+
                threads = NULL;
+               file_hash_exit();
+               fio_debug_jobp = NULL;
+               shmdt(tp);
                shmctl(shm_id, IPC_RMID, &sbuf);
        }
+
+       scleanup();
 }
 
 /*
  * The thread area is shared between the main process and the job
  * threads/processes. So setup a shared memory segment that will hold
- * all the job info.
+ * all the job info. We use the end of the region for keeping track of
+ * open files across jobs, for file sharing.
  */
 static int setup_thread_area(void)
 {
+       void *hash;
+
        /*
         * 1024 is too much on some machines, scale max_jobs if
         * we get a failure that looks like too large a shm segment
@@ -758,6 +860,9 @@ static int setup_thread_area(void)
        do {
                size_t size = max_jobs * sizeof(struct thread_data);
 
+               size += file_hash_size;
+               size += sizeof(unsigned int);
+
                shm_id = shmget(0, size, IPC_CREAT | 0600);
                if (shm_id != -1)
                        break;
@@ -779,6 +884,10 @@ static int setup_thread_area(void)
        }
 
        memset(threads, 0, max_jobs * sizeof(struct thread_data));
+       hash = (void *) threads + max_jobs * sizeof(struct thread_data);
+       fio_debug_jobp = (void *) hash + file_hash_size;
+       *fio_debug_jobp = -1;
+       file_hash_init(hash);
        atexit(free_shm);
        return 0;
 }
@@ -795,12 +904,16 @@ static void usage(const char *name)
        printf("\t--minimal\tMinimal (terse) output\n");
        printf("\t--version\tPrint version info and exit\n");
        printf("\t--help\t\tPrint this page\n");
-       printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
+       printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
+               " them\n");
        printf("\t--showcmd\tTurn a job file into command line options\n");
        printf("\t--eta=when\tWhen ETA estimate should be printed\n");
        printf("\t          \tMay be \"always\", \"never\" or \"auto\"\n");
-       printf("\t--readonly\tTurn on safety read-only checks, preventing writes\n");
+       printf("\t--readonly\tTurn on safety read-only checks, preventing"
+               " writes\n");
        printf("\t--section=name\tOnly run specified section in job file\n");
+       printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
+               " (def 1024)\n");
 }
 
 #ifdef FIO_INC_DEBUG
@@ -812,7 +925,11 @@ struct debug_level debug_levels[] = {
        { .name = "blktrace",   .shift = FD_BLKTRACE },
        { .name = "verify",     .shift = FD_VERIFY },
        { .name = "random",     .shift = FD_RANDOM },
-       { },
+       { .name = "parse",      .shift = FD_PARSE },
+       { .name = "diskutil",   .shift = FD_DISKUTIL },
+       { .name = "job",        .shift = FD_JOB },
+       { .name = "mutex",      .shift = FD_MUTEX },
+       { .name = NULL, },
 };
 
 static int set_debug(const char *string)
@@ -832,22 +949,38 @@ static int set_debug(const char *string)
                }
                log_info("all\n");
                return 1;
-       } else if (!strcmp(string, "all")) {
-               fio_debug = ~0UL;
-               return 0;
        }
 
        while ((opt = strsep(&p, ",")) != NULL) {
                int found = 0;
 
+               if (!strncmp(opt, "all", 3)) {
+                       log_info("fio: set all debug options\n");
+                       fio_debug = ~0UL;
+                       continue;
+               }
+
                for (i = 0; debug_levels[i].name; i++) {
                        dl = &debug_levels[i];
-                       if (!strncmp(opt, dl->name, strlen(opt))) {
+                       found = !strncmp(opt, dl->name, strlen(dl->name));
+                       if (!found)
+                               continue;
+
+                       if (dl->shift == FD_JOB) {
+                               opt = strchr(opt, ':');
+                               if (!opt) {
+                                       log_err("fio: missing job number\n");
+                                       break;
+                               }
+                               opt++;
+                               fio_debug_jobno = atoi(opt);
+                               log_info("fio: set debug jobno %d\n",
+                                                       fio_debug_jobno);
+                       } else {
                                log_info("fio: set debug option %s\n", opt);
-                               found = 1;
                                fio_debug |= (1UL << dl->shift);
-                               break;
                        }
+                       break;
                }
 
                if (!found)
@@ -856,7 +989,7 @@ static int set_debug(const char *string)
        return 0;
 }
 #else
-static void set_debug(const char *string)
+static int set_debug(const char *string)
 {
        log_err("fio: debug tracing not included in build\n");
        return 1;
@@ -868,8 +1001,11 @@ static int parse_cmd_line(int argc, char *argv[])
        struct thread_data *td = NULL;
        int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
 
-       while ((c = getopt_long_only(argc, argv, "", long_options, &lidx)) != -1) {
+       while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
                switch (c) {
+               case 'a':
+                       smalloc_pool_size = atoi(optarg);
+                       break;
                case 't':
                        def_timeout = atoi(optarg);
                        break;
@@ -916,7 +1052,8 @@ static int parse_cmd_line(int argc, char *argv[])
                        break;
                case 'x':
                        if (!strcmp(optarg, "global")) {
-                               log_err("fio: can't use global as only section\n");
+                               log_err("fio: can't use global as only "
+                                       "section\n");
                                do_exit++;
                                exit_val = 1;
                                break;
@@ -926,7 +1063,7 @@ static int parse_cmd_line(int argc, char *argv[])
                        job_section = strdup(optarg);
                        break;
                case FIO_GETOPT_JOB: {
-                       const char *opt = long_options[lidx].name;
+                       const char *opt = l_opts[lidx].name;
                        char *val = optarg;
 
                        if (!strncmp(opt, "name", 4) && td) {
@@ -989,7 +1126,7 @@ int parse_options(int argc, char *argv[])
        f_out = stdout;
        f_err = stderr;
 
-       fio_options_dup_and_init(long_options);
+       fio_options_dup_and_init(l_opts);
 
        if (setup_thread_area())
                return 1;
@@ -1018,5 +1155,11 @@ int parse_options(int argc, char *argv[])
                return 1;
        }
 
+       if (def_thread.o.gtod_offload) {
+               fio_gtod_init();
+               fio_gtod_offload = 1;
+               fio_gtod_cpu = def_thread.o.gtod_cpu;
+       }
+
        return 0;
 }